xAI Exposed for Long-term Distillation of Claude Model Data in Cat-and-Mouse Game
xAI has been conducting a long-term distillation project, training code models directly on outputs from Claude, continuing to extract data through personal accounts and third-party intermediaries like Blackbox AI even after Anthropic banned official channels.
After xAI co-founder Tony Wu reported the official channel was cut off in January, engineers turned to personal accounts and encrypted intermediaries, continuing to obtain data for benchmarking until mid-May; Musk admitted in the OpenAI lawsuit that Grok's training partially used competitor models, stating that cross-model training is a common practice in the industry.
In May, both parties shifted to a monthly $12.5 billion computing power leasing deal, with Anthropic renting xAI's data center to alleviate bottlenecks, and the substantial funds easing xAI's losses while providing an AI growth narrative for SpaceX's IPO.
In market mechanisms, AI labs buy data and computing power as complementary resources, selling purely self-trained dependencies; the cat-and-mouse game exposed by the event and leasing transactions directs funds towards AI infrastructure leasing and model distillation ecosystems, benefiting xAI/SpaceX and flexible data acquirers, while putting pressure on labs with strict data protection and high compliance costs.
Source: Public Information
ABAB AI Insight
xAI has rapidly caught up with leading models like Claude since its inception, and this cat-and-mouse style distillation continues Musk's projects' strategy of leveraging public/semi-public interfaces to accelerate iteration, reflecting a common strategy under data scarcity in early stages, ultimately binding interests through computing power leasing.
In terms of capital, xAI reduces self-training costs through distillation and quickly enhances code model capabilities, motivated to shorten the gap with Anthropic/OpenAI, strategically combining data acquisition with computing power output, alleviating significant losses in Q1, while constructing an AI growth story for SpaceX's IPO next week, and testing the industry's gray boundaries.
Similar to past disputes over data/model sharing among AI labs, xAI is currently transitioning from a follower to an infrastructure provider, with leasing transactions marking a shift from zero-sum data competition to mutually beneficial resource exchange.
Essentially, this represents a restructuring of the industry chain: distillation + computing power leasing replaces completely isolated training with cross-company data flows, breaking through single lab bottlenecks in computing power and data, promoting AI development from closed competition to ecological resource sharing, accelerating the overall industry iteration speed and capital efficiency.
ABAB News · Cognitive Law
When data scarcity meets bans, intermediaries' guerrilla tactics outperform halting iterations.
Competitors are complementary resources; those who lease computing power first can leverage bottlenecks.
The endgame of the cat-and-mouse game is often transactions; those who turn enemies into renters first occupy the pricing power of AI infrastructure.